meet2701/smai-a4-mlp

SMAI A4: MLP on CIFAR-10

Model compared in Section 1.5 of SMAI Assignment 4 (true labels, p = 0, 100 epochs).

• layers: 3072 -> 128 -> 10, hidden activation ReLU
• constraints: none
• init: Normal(0, 2/n_in) for ReLU hidden layers, Normal(0, 1/n_in) for the output layer; biases 0
• training: plain SGD, lr 0.01, batch 64, 100 epochs, float64
• preprocessing: flatten 32x32x3 (H, W, C order) to 3072, divide by 255, then (x - mean) / std with mean/std from preprocessing.npz (train part, per column)
• final test accuracy: 0.4835

Files: model.npz (weights, model.npz: W{i} with shape (n_in, n_out) and b{i}; logits = layer stack applied to x @ W + b), preprocessing.npz (mean, std), config.json.